Corticom Unsupervised Neural VAD

Code repository for the paper "Real-time detection of spoken speech from unlabeled ECoG signals: A pilot study with an ALS participant" by Angrick et al.

Installation

All dependencies can be found in the requirements.txt file. This repository was tested with Python version 3.10. Additionally, the TICC algorithm is linked as a submodule.

pip install -r requirements.txt
pip install extensions/hga
git submodules init
git submodules update

Replicate results

There is a replicate shell script which runs all steps from the paper to reproduce the results. It needs to configured right i the beginning to point to the right paths for accessing the data and storing all results.

Structure of the output folder

The scripts will all output their results in a dedicated folder. This folder will contain the following sub folders:

• analysis: This folder holds all rendered figures after running the replicate shell script.
• corpus: The corpus folder will be created by the prepare_corpus.py script during the computation of the high-gamma features (and normalizations). All results are stored in the HDF5 fle format.
• normalization: This folder will be created by the prepare_corpus.py script. It contains for each particular day the normalization statistics.
• temporal_context: This folder contains the alignment errors (according to the Levenshtein distance) for each trial. It will be created by the compute_temporal_context.py script.
• gen_labels_ticc: Folder containing the estimated labels from the TICC algorithm. Results are stored in the HDF5 file format and contain datasets for the high-gamma features, the acoustic VAD alignments and the alignments from the TICC algorithm. This folder will be created by estimate_vad_labels.py.
• baseline: Results from the baseline computation (Leave-one-day-out cross-validation), split into folders for the CNN and the logistic regression approaches. This folder will be created by baseline_computations.py.